DRAG for AI Upskilling at Manila IT Site

Question (2026-05-30, via Telegram #3097): "Find our DRAG framework from the my second-brain and share with me practical way I can apply in AI upskilling at Manila IT site."

The wiki's canonical source for DRAG is DRAG Framework (concept page), which itself comes from Dangerously Smart with AI (theMITmonk) by Sandeep Swadia. This page is the practical translation — DRAG as the spine of an AI-upskilling program for a corporate IT team, with the Manila site as the deployment context.

What DRAG actually is (in one line)

Sandeep's 4-category delegation rubric: Drafting, Research, Analysis, Grunt work — the "zone 1" (capped payoff) tasks where outsourcing to AI is correct. The decision rule is the cleanest part: apply DRAG only to capped-payoff work; if it requires human judgment / intuition / taste / decision-making, keep it on the human (zone 2).

Sandeep's own estimate: "70 to 80% of my repetitive tasks tend to be in zone one." For a corporate IT team, that band is probably similar — operational tickets, runbook drafting, log triage, status reporting, vendor research.

Why DRAG is the right spine for an IT-team upskilling program

Most enterprise AI-upskilling programs default to tool tours (here's how to prompt ChatGPT, here's how to use Copilot). That gets uptake numbers but not behavior change. DRAG inverts that: it teaches people what to delegate before it teaches them how to delegate. The 30/70 framing (zone 2 stays human, zone 1 goes to AI) is the single most leverage-y mental shift you can install in a non-engineer IT team — once it lands, the tool questions answer themselves.

Three additional reasons DRAG fits an IT-site upskilling rollout specifically:

1. It's binary at the work-item level. People can sort their own backlog into D / R / A / G or "neither" the same week they hear the framework. Cheap to operationalize.
2. It survives without specific tools. Copilot today, Claude tomorrow, OpenClaw next year — the rubric doesn't change. Tool-specific training has a 6-month shelf life; DRAG doesn't.
3. It pairs cleanly with accountability. The "is this zone 1 or zone 2" question is the same one a manager would ask in a review. Builds a shared vocabulary across IC and lead.

A 4-phase rollout playbook

Phase 0 — Pick the cohort and the inventory (week 0)

• Cohort: start with one squad at the Manila site — pref ~8–12 people, mixed levels (IC + lead). Not the whole site at once. Sandeep's own narrow-agents thesis (Narrow Agents / You're Not Behind (Yet) Learn AI Agents (theMITmonk)) is the precedent: narrow first, prove it, then expand.
• Inventory: each person lists their top ~15 recurring work items from the last 30 days. No tools talk yet.
• Pre-test: ask each person to self-classify their list as zone 1 (capped) or zone 2 (uncapped) without DRAG. Keep this. You'll diff it later.

Phase 1 — Install the DRAG vocabulary (week 1)

• 45-min workshop, not a 3-hour course. The cognitive load is one mental model + one decision rule.
• Teach: D / R / A / G + zone-1-vs-zone-2 + the AIM Protocol (Actor / Input / Mission) as the only prompt pattern for drafting tasks. Three things, no more.
• Worked example per category, drawn from the cohort's own inventory (not stock examples). 10 minutes per category.
• Closing exercise: each person reclassifies their inventory using DRAG. Compare to their pre-test. The diff usually surprises them — people underestimate how much of their work is zone 1 because completion bias makes everything feel important (see Dangerously Smart with AI (theMITmonk) §1).

Phase 2 — Supervised practice with the Intelligent Gym pattern (weeks 2–4)

This is where most programs lose people. The fix is to add deliberate friction, not remove it. Sandeep's Intelligent Gym frame: "For information tasks, use AI to remove friction. For transformation tasks, use AI to add friction."

Operationally, for the cohort:

• Two DRAG tasks per week per person, executed using AIM. Submit the prompt, the AI output, and a 1-paragraph reflection on what they edited and why. The edit step is the gym rep — pure-acceptance is the cognitive-offloading trap (see Cognitive Offloading).
• One zone-2 task per week, deliberately not delegated. Annotate why. Builds the discrimination muscle — knowing where not to use AI is half the upskilling.
• One progressive-overload session per fortnight (Sandeep's pattern): pick a technical concept the team is shaky on (e.g., a new cloud service, a security primitive), have AI quiz the engineer at increasing difficulty levels: high-school → college → exec-interview → irate-boss. ~30 minutes. Practical and on-the-job.

The team lead reviews ~20% of submissions weekly — not to judge, to catch Hallucination Laundering (uncritically forwarded AI output) before it becomes a habit.

Phase 3 — Graduate from prompts to agents using ARR (weeks 5–8)

Most IT teams will discover that some of their G-bucket (grunt work) tasks aren't just delegable — they're autonomous, recurring, and reviewable. That's the ARR Framework signal to move from prompt to agent.

For Manila IT specifically, candidates likely include:

• Weekly access-review reports (R + G)
• Ticket triage and routing (A + G)
• On-call rotation summary drafts (D + A)
• Vendor SOC2 / DPA review first-pass (R + A)
• Daily / weekly KPI dashboards (R + G)

For each candidate, run the GPS Check (for Agents) before building anything:

• Goal — can you say what the agent should do in one clear sentence?
• Proof — what does "good output" look like, in measurable terms?
• Steps — can you describe each step without handwaving?

If any of the three fails, the task isn't ready for an agent — it's ready for better SOPs first. That's the right answer too, and arguably the more durable upskilling outcome.

Phase 4 — Measure and expand (week 8+)

Track three numbers per cohort member (not per task):

• Zone-1 displacement rate — % of recurring zone-1 tasks now AI-assisted
• Edit ratio — fraction of AI outputs the person modified meaningfully before using (the inverse of Cognitive Offloading — high is good; >50% is healthy)
• GPS-pass rate — for proposed agents, what fraction passed the GPS Check on first attempt (proxy for clarity of the underlying work, not AI competence)

After 8 weeks, if those numbers look good in cohort 1, expand to cohort 2. Don't expand on cohort 1 results alone — narrow ownership (Narrow Agents) is the same thesis at the program level.

What to push back on / common failure modes

Three traps the framework doesn't surface on its own:

1. DRAG-for-zone-2. People will misclassify and try to delegate strategy / org-design / customer-judgment work into the D bucket. The check is the capped-payoff test: would 1% better effort here actually be worth ~1%, or solve the rest of the 99%? If the latter, it's zone 2 — keep it. (This is the load-bearing point from Sandeep's video; rehearse it explicitly.)
2. Drafting → Hallucination Laundering. The D bucket is the highest-risk one because the polish makes it feel "done." Required discipline: the human edit pass is part of the rep, not optional.
3. Tool sprawl without governance. As cohort 1 succeeds, people will start connecting AI tools to corporate data (Drive, email, calendar). This is exactly the Shadow AI vector. Pair the DRAG rollout with a minimum governance baseline (AWARE Framework is the technical-controls reference; the practitioner-side answer is the Five AI Risks That Can Get You Fired (IBM Technology) checklist). The IT site is the right place to model this because IT-team members are the future implementers of AI governance at the wider company.

The wider through-line — why this is on-thesis for IT leadership

DRAG is the personal-productivity layer of a 3-layer stack the vault already has covered:

• Personal productivity (this page) — DRAG, AIM, Intelligent Gym
• Team-architecture — Bounded vs Unbounded Tasks (Praveen's enterprise version of the same decision), Build vs Buy (Agents), ARR Framework
• Governance — AWARE Framework, CIO Agenda 2026 (CXOTalk), Five AI Risks That Can Get You Fired (IBM Technology)

Running the DRAG-upskilling program at one IT site is a demonstration project for the team-architecture and governance layers at the wider org. The Manila cohort isn't just learning to use AI — they're piloting the operating model the rest of the company will need within ~12 months. That framing also makes the program defensible to a CFO: it's not L&D spend, it's an early-investment in the org's Code Is Free / judgment-scarce future.

Brand fodder

The Medium / LinkedIn version of this page writes itself. Suggested framing for the user's senior-IT-leader audience: "We don't have an AI tools problem. We have a delegation problem. Here's the rubric we installed at one IT site to fix it." Lead with the DRAG mental model (zone 1 vs zone 2), tell the Manila pilot story, close on "narrow ownership beats broad coverage." On-thesis because it reframes AI upskilling as operational discipline, not training.

Cross-links

• Concepts · DRAG Framework · AIM Protocol · Intelligent Gym · GPS Check (for Agents) · ARR Framework · Bounded vs Unbounded Tasks · Cognitive Offloading · Hallucination Laundering · Shadow AI · AWARE Framework · Narrow Agents · Code Is Free
• People · Sandeep Swadia
• Sources · Dangerously Smart with AI (theMITmonk) · You're Not Behind (Yet) Learn AI Agents (theMITmonk) · Five AI Risks That Can Get You Fired (IBM Technology) · CIO Agenda 2026 (CXOTalk)

Source trigger

• Telegram message #3097 (2026-05-30 13:43) — captured in raw/processed/